Dynamics of localized & distributed codes across regions of primate frontal cortex during multi-step inference

Jascha Achterberg*, Valentina Mione*, Mark J. Buckley, Daniel Mitchell, and John Duncan

This is a current summary of results. We are currently preparing a preprint. Please reach out if you are interested in more details.

Background & question

Frontal cortex (FC) allows for the complex problem- solving abilities observed in primates. Prior recordings in FC were either focused on simple tasks or isolated regions.

Using novel large-scale electrophysiology recordings in monkey FC we want to understand how subregions of FC come together to jointly orchestrate complex computations.

Task & recordings

Monkeys solve a multi-step maze task which requires them to navigate a 2D grid from the start location to 1 of 4 goal locations, using saccades. Monkeys start at the center, need to remember the current goal location (presented at the start of the trial), and navigate using presented choice options:

Depending on available choice options, goal locations can be reached via 2-step or 4-step routes:

We record 2865 neurons from 6 regions of frontal cortex, using semi-chronic microelectrode arrays:

Extracting neural subspaces

To understand the computations within FC, we want to identify the neural subspaces in which key variables are represented (goal, next move):

Combine data from 2nd step & 4th step to create PCA space.
Project data from across the trial into space to observe the dynamics of the geometry.

For each region we ask: Which variables determine the shape & dynamics of projections?

Codes for goal and location

We project the population activity for trials into the ‘Goal space’ and then measure the average distance of projections when grouping them by ‘current goal’ or ‘current position’.

Illustration: Grouped by shape and color. Distances lower for color.

Then we compare this distance to cases with different goals & positions (‘Baseline’). Data from 2nd & 3rd choice. P-values via bootstrap. (Not depicted: dmPFC has stable goal code throughout trial.)

Codes for next move

Same as (4) but using the ‘Move space’ and calculating distances with regards to the first move or second move. ‘Baseline’ are moves which are neither first nor second move. All regions show significant move code.

Generally, the ‘Move space’ is orthogonal to the ‘Goal space’, except for in dmPFC.

Hierarchical choice code

When comparing the similarity of representations (‘population activation vectors’) across time without PCA filtering, regions show unique hierarchical structure. Not depicted regions look like In/OFC.

Conclusions

Regions of PFC jointly solve the complex multistep problem. Individual regions have their preferred variables, synchronizing computations with weaker copies of other variables. Jointly they represent goals, current state, anticipated policy, and the overall task hierarchy.